Module 01

Intelligence that can change providers without changing the product.

The LLM layer helps with ambiguous product ranking, basket suggestions, explanations, and recovery hints. Structured contracts keep model output advisory and testable.

2provider adapters

5initial task types

1shared evaluation set

0purchase tools

Flow

Provider-neutral orchestration

Intelligence module — Task policy selects provider, model, prompt, schema, timeout, and fallback. All outputs are normalized and schema-validated before returning to the workflow.

Contract

Stable input and output

interface IntelligencePort {
  generateStructured<TInput, TOutput>(
    request: IntelligenceRequest<TInput>,
    outputSchema: JsonSchema
  ): Promise<IntelligenceResult<TOutput>>;
}

type IntelligenceTask =
  | 'basket.propose'
  | 'product.match-candidates'
  | 'substitution.rank'
  | 'order.explain-changes'
  | 'browser.recovery-hint';

Task policy

Use models only where they add value

Task	Model contribution	Deterministic verification
Basket proposal	Suggest likely due items from sanitized history features	User confirmation, cadence rules, active product refs
Candidate ranking	Rank a small set of similar retailer products	Exact ID, size, availability, approved policy, fresh price
Substitution ranking	Explain tradeoffs among already allowed candidates	Approved alternatives and item-level limits
Change explanation	Turn reason codes into concise household language	Every statement traces to structured facts
Recovery hint	Suggest how to re-locate a changed page element	No direct action; reviewed locator and postcondition required

Adapters

OpenAI and Anthropic behind one port

Primary or fallback

OpenAI adapter

Responses API with a configured GPT-5.6 model, structured output, normalized usage, bounded timeout, and task-specific prompt version.

Primary or fallback

Anthropic adapter

Messages API with claude-opus-4-8, equivalent schemas, normalized finish reasons, and the same contract suite.

Provider swap is an evaluation event

Changing the configured model is not only a configuration edit. The candidate-matching, injection, faithfulness, schema-validity, latency, and cost evaluation set must pass again.

Safety

What remains outside the model

Authoritative totals

Prices, fees, quantities, savings, and budget enforcement come from retailer data and deterministic calculation.

Product identity

A generated product reference is never trusted. It must resolve against a fresh retailer candidate and pass product rules.

Purchase authority

Automation mode, approval validity, restricted categories, and final checkout are evaluated outside the LLM.